Microsoft AB-100 (Agentic AI Business Solutions Architect) Practice Test 110

Question 1090

Security and business owners want to keep this capability: Answers depend on policies revised every week. They must also prevent the following: A model might state outdated rules as facts when authoritative policies have changed. Which guardrail should the design incorporate?

  1. Disallow factual claims unsupported by current policy data
  2. Use approved extension points rather than unsupported modifications
  3. Promote complex or low-confidence cases to stronger fallback
  4. Add human approval for consequential recommendations

Explanation

Correct Answer

A:

When a model might state outdated rules as facts when authoritative policies have changed, the use case answers depend on policies revised every week requires a guardrail aligned with the behavior that retrieval separates volatile knowledge from model parameters.

Incorrect Answers

B:

Another process could need this guardrail because an unsupported product customization might break after vendor updates; the design being assessed supports answers depend on policies revised every week and has the explicit exposure that a model might state outdated rules as facts when authoritative policies have changed.

C:

The safeguard for many calls require only a small fixed-label output deals with a different threat: a small model might make a high-complexity judgment with insufficient confidence; even after it is enforced, the exposure that a model might state outdated rules as facts when authoritative policies have changed remains possible in the target workflow where retrieval separates volatile knowledge from model parameters.

D:

Another process could need this guardrail because an automated recommendation might be executed without human accountability; the design being assessed supports answers depend on policies revised every week and has the explicit exposure that a model might state outdated rules as facts when authoritative policies have changed.

Question 1091

A risk analysis highlights a threat in this workflow: Agents modify records through external APIs. Exposure: Retries might cause repeated transactions even though tool calls appear successful. Which enforceable safeguard addresses it at the relevant boundary?

  1. Use idempotency on any repeatable mutation action
  2. Fail closed on authorization and schema-validation failures
  3. Reject requests independent of instruction wording
  4. Require mitigation before deployment when disparity is material

Explanation

Correct Answer

A:

Avoiding the consequence where retries might cause repeated transactions even though tool calls appear successful requires a control for agents modify records through external APIs; in this architecture, tool metrics reveal mutation failures and retry consequences, which is the point at which the boundary can be applied.

Incorrect Answers

B:

Another process could need this guardrail because malformed calls might pass a happy-path suite yet mutate records; the design being assessed supports agents modify records through external APIs and has the explicit exposure that retries might cause repeated transactions even though tool calls appear successful.

C:

Where a user asks the agent to read an unrelated executive record, the guard can address a persuasive jailbreak might cause the backend to reveal privileged records; the exposure that retries might cause repeated transactions even though tool calls appear successful remains relevant to the present assessment.

D:

That protection is tied to the mechanism that segmented evaluations reveal uneven model behavior; it does not prevent the distinct risk where retries might cause repeated transactions even though tool calls appear successful in the system under review.

Question 1092

The acceptance board asks how to verify a corrected design. The setting is: A canvas app passes customer records to a model action. Last time, hidden fields from the UI appear in the prompt payload. Which scenario should the team execute?

  1. Query as multiple roles and confirm only authorized fields appear
  2. Try the action with a user lacking record access
  3. Simulate failed verification and confirm no notification
  4. Open the page for different roles and verify context limits

Explanation

Correct Answer

B:

The check is targeted rather than generic: server-side authorization and minimized inputs protect data flow; its pass/fail evidence applies to the workflow where a canvas app passes customer records to a model action and to the prior defect that hidden fields from the UI appear in the prompt payload.

Incorrect Answers

A:

This check exercises a different system behavior: field projection and endpoint authorization minimize exposed data; a pass there cannot establish that hidden fields from the UI appear in the prompt payload no longer occurs in the required workflow.

C:

For approval processing requires sequential verification and notification, the alternate test is about whether defined flow steps make sequencing and error paths explicit; it cannot prove the outcome needed after the reported regression where hidden fields from the UI appear in the prompt payload.

D:

Although the exercise might catch generic chat loses visibility into related records and actions, it would not directly validate the mechanism that server-side authorization and minimized inputs protect data flow in the target implementation.

Question 1093

A design adds autonomous behavior to a process where some cases need cross-document reasoning and detailed justification. This creates a documented risk: An automated recommendation might be executed without human accountability. Which enforced control is the best fit?

  1. Disallow factual claims unsupported by current policy data
  2. Add human approval for consequential recommendations
  3. Budget for independent hosting, safety testing and monitoring
  4. Require risk sign-off from accountable stakeholders

Explanation

Correct Answer

B:

Avoiding the consequence where an automated recommendation might be executed without human accountability requires a control for some cases need cross-document reasoning and detailed justification; in this architecture, complexity classification selects a capable model for difficult tasks, which is the point at which the boundary can be applied.

Incorrect Answers

A:

The safeguard for answers depend on policies revised every week deals with a different threat: a model might state outdated rules as facts when authoritative policies have changed; even after it is enforced, the exposure that an automated recommendation might be executed without human accountability remains possible in the target workflow where complexity classification selects a capable model for difficult tasks.

C:

The alternate boundary mitigates a bespoke runtime might be shipped without an accountable security or support team, which is relevant to a workflow needs proprietary scheduling, model choice and runtime control; it leaves the current workflow vulnerable to an automated recommendation might be executed without human accountability.

D:

Where business, platform, security and legal teams make conflicting AI decisions, the guard can address cross-function objections might be overruled without an accountable approval decision; the exposure that an automated recommendation might be executed without human accountability remains relevant to the present assessment.

Question 1094

Release reviewers will not accept a generic success message. The business scenario is that prompt revisions cause intermittent regressions, and the historical failure is: Makers validate only a few manual chats. Which test is appropriately targeted?

  1. Evaluate candidate and baseline on identical frozen examples
  2. Simulate elevated error rates and verify alert routing
  3. Run the same corpus in a CI release candidate
  4. Create invalid ID, wrong-tenant and malformed-amount examples

Explanation

Correct Answer

C:

The scenario is that prompt revisions cause intermittent regressions; the defect to rule out is that makers validate only a few manual chats; testing via running the same corpus in a CI release candidate challenges the design dependency where stable automated suites compare response quality across versions.

Incorrect Answers

A:

For teams promote a custom model because it feels better, the alternate test is about whether predefined gates weigh accuracy, latency and business risk; it cannot prove the outcome needed after the reported regression where makers validate only a few manual chats.

B:

The tested behavior is that service-level objectives and alert thresholds turn telemetry into response signals; the necessary evidence here concerns stable automated suites compare response quality across versions and the previously observed makers validate only a few manual chats.

D:

Creating invalid id, wrong-tenant and malformed-amount examples verifies an adjacent objective where agents can call APIs with customer IDs and amount values; it leaves untested the causal path in which stable automated suites compare response quality across versions for this case.

Question 1095

Before introducing a fix, the team must explain why the agent confidently answers when retrieval fails. The affected setting is: Policy answers must be tied to a current source. What should be inspected?

  1. Compare answers with retrieved and deliberately missing documents
  2. Compare retrieved document IDs for multiple test principals
  3. Compare requested export destinations against policy allowlists
  4. Compare errors by relevant justified cohorts with privacy care

Explanation

Correct Answer

A:

For the situation where policy answers must be tied to a current source, the defect is that the agent confidently answers when retrieval fails; comparing answers with retrieved and deliberately missing documents helps verify the cause-and-effect path: evidence and refusal evaluations enforce source-supported claims.

Incorrect Answers

B:

The proposed inspection is associated with users have different rights to the same knowledge corpus, where retrieval applies effective source entitlements to candidate documents; it would not expose the failure mode that the agent confidently answers when retrieval fails with the required specificity.

C:

In the separate scenario where a prompt injection requests data export to an outside address, the same investigation could explain model-level refusal alone may not prevent a tool call; it does not isolate the reported cause in a system where policy answers must be tied to a current source.

D:

Overall accuracy conceals disproportionate error; examining that other path does not determine whether evidence and refusal evaluations enforce source-supported claims, which is what distinguishes the issue in the stated use case.

Question 1096

Troubleshooting starts from a business complaint: The agent lacks knowledge of non-Microsoft content. The affected process is: Teams want approved external records searchable in Copilot. Which investigation is most directly tied to the reported defect?

  1. Inspect connector crawl scope, schema and ACL mappings
  2. Document specific requirements unmet by available platforms
  3. Inspect routing decisions and specialist task descriptions
  4. Inspect pending tasks and final-result callbacks by correlation ID

Explanation

Correct Answer

A:

For the situation where teams want approved external records searchable in Copilot, the defect is that the agent lacks knowledge of non-Microsoft content; inspecting connector crawl scope, schema and ACL mappings helps verify the cause-and-effect path: connectors expose external enterprise content to supported Microsoft search/Copilot experiences.

Incorrect Answers

B:

In the separate scenario where competitive requirements depend on unique multi-model control, the same investigation could explain a boxed product cannot express the execution and ownership rules; it does not isolate the reported cause in a system where teams want approved external records searchable in Copilot.

C:

The proposed probe concerns a request spans procurement, finance and policy expertise, where a coordinating agent selects bounded specialists and merges their outputs; this incident instead requires evidence about connectors expose external enterprise content to supported Microsoft search/Copilot experiences and the failure that the agent lacks knowledge of non-Microsoft content.

D:

Inspecting pending tasks and final-result callbacks by correlation id explores a separate connection, where task IDs, status transitions, deadlines and durable handoff state support coordination; the diagnosed requirement is teams want approved external records searchable in Copilot, and the actual defect is the agent lacks knowledge of non-Microsoft content.

Question 1097

A quality engineer is writing a regression case after the incident where agent sends incomplete requests when users answer out of order. The underlying business scenario is: An intake task requires several prerequisite fields. What should the test exercise?

  1. Simulate failed verification and confirm no notification
  2. Generate stressful and adversarial user examples
  3. Force a downstream failure after the initial write
  4. Supply fields in unusual order and verify completeness

Explanation

Correct Answer

D:

The historical failure is that agent sends incomplete requests when users answer out of order; for the setting where an intake task requires several prerequisite fields, supplying fields in unusual order and verifying completeness verifies the dependency where state tracking captures required inputs before any action.

Incorrect Answers

A:

Simulating failed verification and confirming no notification verifies an adjacent objective where approval processing requires sequential verification and notification; it leaves untested the causal path in which state tracking captures required inputs before any action for this case.

B:

A pass on that verification would speak to compliance letters must preserve approved language and disclaimers; it would not show that state tracking captures required inputs before any action, which is needed to avoid the observed defect agent sends incomplete requests when users answer out of order.

C:

Forcing a downstream failure after the initial write verifies an adjacent objective where an agent creates an order but reservation of inventory fails; it leaves untested the causal path in which state tracking captures required inputs before any action for this case.

Question 1098

The architecture board has evidence that retrieval mixes records from two organizations. The affected situation is: Many tenants have identically named projects. What change most directly establishes the missing behavior?

  1. Use hybrid semantic and lexical retrieval for specialist terms
  2. Apply entity-aware filtering before grounding customer answers
  3. Choose document intelligence extraction for complex business forms
  4. Use managed inference and operational health checks for a custom model
  5. Version models and datasets together for controlled adaptation
  6. Extend service assistance with an approved escalation action

Explanation

Correct Answer

B:

Because tenant and entity filters constrain candidate records before generation, this approach corrects the reported breakdown where retrieval mixes records from two organizations; the affected situation is that many tenants have identically named projects.

Incorrect Answers

A:

When product catalog queries mix synonyms and exact serial codes, hybrid ranking combines lexical precision and semantic relevance; in the stated workflow, however, retrieval mixes records from two organizations; the relevant architectural behavior is that tenant and entity filters constrain candidate records before generation.

C:

There is a valid place for the mechanism that document processing models extract structured fields from document content, namely vendors send variable-layout invoices and shipping documents; it does not remove the specific defect that retrieval mixes records from two organizations from this case.

D:

This design would make sense if a specialized model must serve production apps reliably, particularly where model quality is strong but endpoint latency breaks the workflow; the requested architecture instead needs the dependency that tenant and entity filters constrain candidate records before generation because retrieval mixes records from two organizations.

E:

Dataset lineage and model versioning support repeatable assessment; its main purpose is retraining uses revised labels and a different source snapshot, not correction of the reported defect where retrieval mixes records from two organizations in the target process.

F:

An authorized action creates a governed workflow item from the current case; its main purpose is a representative must initiate an exception workflow from the case, not correction of the reported defect where retrieval mixes records from two organizations in the target process.

Question 1099

Operations has not established the cause of this defect: A separate application breaks established productivity workflows. The associated setting is: Staff need a role-specific conversation within Copilot and Teams. Which inspection should precede remediation?

  1. Inspect routing decisions and specialist task descriptions
  2. Compare gap size to supported extension interfaces
  3. Document specific requirements unmet by available platforms
  4. Verify required host and distribution capabilities

Explanation

Correct Answer

D:

The reported fault is that a separate application breaks established productivity workflows in a system where staff need a role-specific conversation within Copilot and Teams; evidence from verifying required host and distribution capabilities can show whether declarative agents use host orchestration with tailored instructions and actions.

Incorrect Answers

A:

In the separate scenario where a request spans procurement, finance and policy expertise, the same investigation could explain one general agent conflates expertise and may call the wrong tools; it does not isolate the reported cause in a system where staff need a role-specific conversation within Copilot and Teams.

B:

The proposed probe concerns current users need one governed custom action inside their workflow, where extension retains familiar user experience and adds bounded custom behavior; this incident instead requires evidence about declarative agents use host orchestration with tailored instructions and actions and the failure that a separate application breaks established productivity workflows.

C:

In the separate scenario where competitive requirements depend on unique multi-model control, the same investigation could explain a boxed product cannot express the execution and ownership rules; it does not isolate the reported cause in a system where staff need a role-specific conversation within Copilot and Teams.

Question 1100

A risk analysis highlights a threat in this workflow: An AI-driven replenishment proposal influences procurement. Exposure: Autonomous replenishment proposals might place purchases beyond approved limits. Which enforceable safeguard addresses it at the relevant boundary?

  1. Do not allow autonomous purchasing without approved limits
  2. Stage compatible transitions and protect authoritative fields
  3. Do not approve procurement while financial hold status is unknown
  4. Prevent sensitive data from entering prompts unless required

Explanation

Correct Answer

A:

When autonomous replenishment proposals might place purchases beyond approved limits, the use case an AI-driven replenishment proposal influences procurement requires a guardrail aligned with the behavior that a pilot cohort and rollback path isolate operational impact.

Incorrect Answers

B:

The alternate boundary mitigates cross-Dynamics releases might overwrite authoritative customer fields, which is relevant to a customer handoff depends on shared record and term mappings; it leaves the current workflow vulnerable to autonomous replenishment proposals might place purchases beyond approved limits.

C:

Another process could need this guardrail because a procurement approval might proceed before a finance hold is resolved; the design being assessed supports an AI-driven replenishment proposal influences procurement and has the explicit exposure that autonomous replenishment proposals might place purchases beyond approved limits.

D:

The proposed control targets a model might expose sensitive user data in uncontrolled prompt output, while the given risk scenario is autonomous replenishment proposals might place purchases beyond approved limits; these exposures arise in different operational contexts.

Agentic AI Business Solutions Architect (AB-100) Practice Tests